Skip to content

Landscape survey: open-source EM scattering solvers (Mie, T-matrix, DDA/GDM, FDTD)

Compiled 2026-11-01 for Lightsmith, an MIT-core, uv/pip-installable Python library with CPU and GPU support.

How the facts were gathered

Tag Source and method
[PyPI] PyPI JSON API (https://pypi.org/pypi/<pkg>/json), queried 2026-11-01 for latest version, upload date, license field, wheel tags and requires_dist.
[GH] GitHub REST API: pushed_at, SPDX license, stars, latest release.
[GL] GitLab REST API, the same fields.
[doc] The project's own README, docs or paper page, fetched on the survey date.
unverified Could not be confirmed from a primary source.

Wheel tags. win means a native Windows win_amd64 wheel. any means a pure-Python wheel, which installs anywhere, including with uv. sdist means a source build; for Fortran or C++ packages that needs a local compiler, gfortran in particular, which is painful on Windows.

uv. Anything on PyPI with a matching wheel installs with uv. The caveats are PyTorch CUDA builds, which need the PyTorch package index, and conda-only packages.

0. Executive summary

  1. No direct competitor exists. No maintained package puts Mie, T-matrix, DDA and FDTD behind one problem spec. The closest partial overlaps:
  2. HoloPy (GPL-3) wraps Mie, multisphere T-matrix and ADDA, but it targets holography, and its last release was 2021.
  3. TorchGDM (GPL-3) combines GDM with Mie and T-matrix effective models via treams.
  4. Tidy3D is cloud-only for its solver.
  5. dda-bench (GPL-3, 2026) cross-verifies three DDA codes only.
  6. A permissive GPU + autodiff FFT-DDA is a real gap.
  7. TorchGDM: GPL-3, dense matrices, about 10–15k dipoles, homogeneous environment only.
  8. pyGDM2: GPL-3. It uses CuPy only for dense inversion.
  9. CPDDA has CuPy FFT-DDA, but no stated license, no PyPI package and no autodiff.
  10. ADDA and DDSCAT are GPL executables.
  11. The GPU-strategy decider: JAX still has no official native-Windows CUDA support (late 2026). The jax-cuda12/13-plugin wheels are Linux-only, and WSL2 GPU is "experimental". The only native route is the unofficial winjax beta. That beta is single-GPU and cannot share a process with PyTorch.
  12. PyTorch ships native Windows CUDA wheels; torch 2.15.0 has cu132 and cu134 builds. PyTorch is therefore the only mainstream autodiff GPU stack that works on the Windows dev machine.
  13. CuPy also has Windows wheels, but it has no autodiff.
  14. Several "standard" backends have packaging rot. These are good upstream-contribution targets.
  15. smuthi (MIT) hard-pins numpy<2 and scipy<1.14. It has a hard dependency on GPL-3 pywigxjpf, and it ships wheels only for CPython 3.8–3.10.
  16. pytmatrix cannot build on Python ≥3.12 because it uses numpy.distutils.
  17. scattnlay is v2.4 on PyPI (cp311/cp312, Windows and macOS-arm wheels only), while the README advertises v2.5.
  18. Meep is still conda-only, with no native Windows support.
  19. pyMSTM has Linux wheels only.
  20. FDTDX (MIT, JAX) gained Drude/Lorentz dispersive materials only in May 2026 (v0.6.2). It has no published plasmonic validation, so our Mie-referenced benchmark would be a genuine service to that community.

1. Mie theory (sphere and layered-sphere ground truth)

Library License Latest release Activity Wheels / uv GPU Autodiff Computes Notes and limitations Citation
miepython (GH, docs) MIT [PyPI] 3.3.1, 2026-10-28 [PyPI] Very active; push 2026-10-28 [GH] any, pure Python plus optional numba; uv OK on all OSes; Python ≥3.11 No No Q_ext, Q_sca, Q_back, g, angular intensities, amplitude functions, Mueller matrices, near-field E/H [doc] Homogeneous spheres only; no coated or layered spheres listed [doc]. Numba JIT gives 10–50× [doc]. Prahl, Zenodo DOI 10.5281/zenodo.7949263 [doc]
scattnlay (GH) GPL-3 [GH] PyPI 2.4, 2024-04-10. The README cites 2.5 (Dec 2025), which is not on PyPI [PyPI], [doc]. Moderate; push 2026-01 [GH] Win and macOS-arm wheels for cp311/cp312 only; Linux and other Pythons need an sdist (C++) build No No Multilayer spheres; near and internal fields; optional Boost multiprecision; SIMD [doc] GPL, so optional plugin only. The best reference for multilayer and near fields; arbitrary precision is useful for a ground-truth audit. Peña & Pal, CPC 180, 2348 (2009); Ladutenko et al., CPC 214, 225 (2017) [doc]
PyMieScatt (GH) MIT 1.8.1.1, 2020-12-30 [PyPI] GitHub push 2025-09, but no release since 2020 sdist only, pure Python, so uv builds it No No Homogeneous and core-shell spheres; inverse Mie; size distributions Stale on PyPI. The Tidy3D docs use it as their Mie reference. Sumlin et al., JQSRT 205, 127 (2018) (citation from memory, unverified)
PyMieSim (GH) MIT [PyPI] 6.1.6, 2026-10-27 [PyPI] Very active, about 120 releases Win, Linux-x86_64 and macOS-arm wheels, cp311–cp314 No No Spheres, core-shell, infinite cylinders; cross sections, far field, detector coupling, near-field example [doc] Hard-pins numpy==2.2.6, matplotlib==3.10.9 and pandas [PyPI requires_dist], which is an environment-conflict hazard as a dependency. Poinsinet de Sivry-Houle et al., Opt. Continuum 2(3), 520 (2023), doi:10.1364/OPTCON.473102 [doc]
PyMieDiff (pymiediff, GH) GPL-3 [PyPI], [GH] 0.13, 2026-10-06 [PyPI] Active (new in 2025) any; Linux and Windows, Python 3.10–3.13 [doc] Yes, PyTorch (CUDA) Yes, torch autograd Multilayer spheres; Q_sca, Q_ext; angular scattering; internal and near fields; refractiveindex.info YAML [doc] GPL, so optional plugin. GPU pays off only for thousands of concurrent evaluations [doc]. APL Photonics 11, 046114 (2026), doi:10.1063/5.0316814 [doc]; arXiv:2512.08614
treams sphere solver (GH) MIT 0.4.7, 2026-07-16 [PyPI] Active; push 2026-10-16 [GH] Win, Linux-x86_64 and macOS-arm wheels, cp39–cp313; <3.14; pins scipy<1.17 [PyPI] No No Sphere T-matrix in helicity or parity basis; multilayer spheres via TMatrix.sphere with several radii (from memory, unverified) Gives a T-matrix-native Mie for consistency checks against the T-matrix backend. Beutel, Fernandez-Corbaton, Rockstuhl, CPC 297, 109076 (2024) [doc]
pymiecs (GL) GPL-3+ [PyPI] 0.5, 2025-12-16 Low activity sdist (pure Python) No No Core-shell spheres Used as a test dependency of pyGDM2 and TorchGDM. —
MiePy (miepy, GH) GPL-3 [GH] 1.1.0, 2026-03-18 [PyPI] Revived in 2026 Win, Linux and macOS (x86 and arm) wheels, cp311–cp313 No No GMMT clusters, i.e. multi-sphere T-matrix; also non-spherical particles via EBCM Strictly a cluster T-matrix code (see §2). GPL. —

Takeaway. Mie is cheap, so it needs no GPU. The CPU ground truth should be miepython (MIT) for homogeneous spheres and treams (MIT) for layered spheres and T-matrix consistency. scattnlay (GPL) is useful as an optional high-precision cross-check. An MIT, PyTorch-batched, differentiable Mie is a small piece of code that does not exist permissively today. PyMieDiff is GPL, so writing our own is a cheap win.

2. T-matrix / multiple-scattering codes

Library License Latest release Activity Wheels / uv GPU Autodiff Computes Notes and limitations Citation
treams (GH, docs) MIT 0.4.7, 2026-07-16 Active (44 stars, 42 open issues) Win, Linux-x86_64 and macOS-arm wheels, cp39–cp313 No [doc] No [doc] Spherical- and cylindrical-wave T-matrices; clusters; 1D, 2D and 3D lattices; stratified media (S-matrices); band structures; helicity and parity bases; chiral materials [doc] No arbitrary-shape (EBCM) T-matrix generator. Pins python<3.14 and scipy<1.17. Same group (KIT) as the T-matrix HDF5 format and database. CPC 297, 109076 (2024)
smuthi (GL, docs) MIT (but see notes) 2.2.4, 2025-10-23 [PyPI] Active on GitLab (2026-10-30) [GL]; master moving to a Meson build 2.2.4 wheels are Linux cp38–cp310 only. 2.2.2 and 2.2.3 had Windows cp38–cp310 wheels. Anything newer means an sdist build with gfortran [PyPI]. Optional PyCUDA (extra cuda), Linux-oriented No Multiple particles near or between planar interfaces; spheres, spheroids, cylinders and polyhedra via NFM-DS (Fortran); near and far fields; cross sections [doc] Hard dependencies: numpy<2.0, scipy<1.14, and pywigxjpf (GPL-3). Windows uses pywigxjpf-win, which only has cp36–cp39 wheels. Not installable cleanly in a modern Python 3.12 or 3.13 environment. Egel et al., JQSRT 273, 107846 (2021), arXiv:2105.04259
pytmatrix (GH) MIT 0.3.3, 2023-05-10 Dormant sdist zip only. setup.py uses numpy.distutils [GH source], so it does not build on Python ≥3.12. No No Mishchenko EBCM for spheroids, cylinders and Chebyshev particles; orientation averaging; radar quantities Effectively uninstallable on modern Python without a Meson port. Leinonen, Opt. Express 22, 1655 (2014) (from memory, unverified)
MSTM v4 (GH) and pyMSTM (GH) MSTM: MIT [GH]. pyMSTM: MIT, with bundled GCC runtime (GPL-3 + GCC exception, LGPL) [PyPI] MSTM tag v4.0.0 (push 2024-02). pyMSTM 2.1.0, 2026-10-15. MSTM slow; pyMSTM new (2026, 0 stars) pyMSTM: Linux-x86_64 wheels only, cp311–cp314; other OSes build from sdist with gfortran [doc]. MSTM itself ships Windows .bat build scripts and MS-MPI. No No Sphere clusters, spheres in layered media, periodic lattices, random orientation, T-matrices, Mueller matrices [pyMSTM README] Spheres only. The MSTM v4 paper was not located (unverified). Mackowski & Mishchenko, JQSRT 112, 2182 (2011)
TERMS (GH) MPL-2.0 [GH] (file-level copyleft; can be used as an external executable) v1.2.0, 2026-05-05 [GH] Low but alive Fortran source; not on PyPI No No Superposition T-matrix for particle clusters; near and far field; orientation-averaged cross sections Not a Python package; it would need a subprocess wrapper. Schebarchov et al., arXiv:2111.12556 (user guide)
SMARTIES (GH) CC BY-NC 4.0, not OSI [search; GH says NOASSERTION] v1.2.0, 2024-07 Dormant MATLAB No No Spheroids, with high accuracy at large aspect ratios Exclude: the non-commercial license is incompatible. Use it only as a published reference. Somerville, Auguié, Le Ru, JQSRT 174, 39 (2016); arXiv:1511.00798
CELES (GH) BSD-3 [doc] v2.2, 2021-04-13 Stale (push 2023-06) MATLAB + CUDA MEX CUDA, via MATLAB; tested on Windows and Linux No Very large sphere ensembles (10k+ spheres) Requires MATLAB, so not pip-installable. The only GPU T-matrix of note. Egel et al., JQSRT 199, 103 (2017) [doc]
MiePy GPL-3 1.1.0, 2026-03 Revived Wheels on all OSes No No GMMT clusters, EBCM non-spherical particles, optical forces GPL, so optional plugin. —
ScatterPy GPL-3+ 0.1.1, 2020 Dead any — — T-matrix Ignore. —

Community infrastructure. - T-matrix HDF5 data format: Asadova et al., JQSRT 333, 109310 (2025), arXiv:2408.10727. - T-matrix database ("Daphona"): arXiv:2602.02101, tmatrix.scc.kit.edu.

We should read and write the HDF5 format natively. It is the obvious interchange format between our T-matrix, DDA and FDTD backends: compute a T-matrix with DDA or FDTD, then compare it against treams.

GPU / autodiff T-matrix. No maintained, pip-installable GPU or differentiable T-matrix code was found. A 2024 paper built a JAX T-matrix model for time-varying nanostructures (arXiv:2409.04551), but no package was released (unverified). This is a secondary gap: T-matrix cost is usually dominated by small dense solves, so the CPU is fine.

3. DDA / coupled-dipole / Green dyadic method

Library License Latest release Activity Install GPU Autodiff Computes Notes and limitations Citation
ADDA (GH) GPL-3 v1.4.0, 2020-12-15 (asset adda-1.4.0-win64.zip) [GH] Repo still active (push 2026-03-30, 157 open issues) Prebuilt Windows binaries; build from source on Linux and macOS; not on PyPI (the PyPI adda package is unrelated) OpenCL (works on Windows) [doc] No FFT-accelerated DDA; MPI; particles near a substrate; orientation averaging; many polarizability prescriptions (LDR, FCD, IGT); near fields Accuracy reference for DDA. The executable can be called via subprocess as an optional plugin. Yurkin & Hoekstra, JQSRT 112, 2234 (2011) [doc]
DDSCAT (page) GPL (v2 or v3 is not clear; unverified) 7.3.3 per Draine's page; date unverified Effectively frozen Fortran source tarball No No DDA with FFT; periodic targets (1D and 2D); orientation averaging; near fields The historic standard. ScatPy (GPL-3, 2014) is a dead Python wrapper. Draine & Flatau user guide, arXiv:1305.6497
IFDDA (page) Unverified Unverified Active (used in the 2026 dda-bench paper) Fortran with a Qt GUI; hosted on GitLab GPU per the dda-bench paper (unverified for IFDDA itself) No DDA; IF-DDAM variant for layered substrates — Chaumet et al. (see dda-bench)
pyGDM2 (GL) GPL-3+ 1.1.12, 2025-11-20 [PyPI] Moderate (2026-04) [GL] any (numba-based now); uv OK on all OSes; pins numpy<=2.3 Optional CuPy dense matrix inversion (Windows OK) No GDM / coupled dipole; substrate via mirror-dipole approximation; 2D and 3D; near and far fields; spectra; EELS/CL; evolutionary optimization Dense matrices, so N is limited to around 10⁴. Wiecha, CPC 233, 167 (2018) (from memory); arXiv:1802.04071
TorchGDM (GL, paper) GPL-3+ [PyPI], [GL] 0.61, 2026-10-26 [PyPI] (GitLab tag v0.62) Very active any; Linux, Windows and macOS; torch>=2.6 PyTorch (CUDA, so native Windows works) Yes, full autograd (except the treams and T-matrix import paths) 3D and 2D GDM; mixes full discretization with effective electric/magnetic dipole (GPM) models; cross sections, near and far fields, LDOS, multipoles, chirality [paper] Dense full inversion, no FFT or iterative solver; memory grows as N²; about 10k–15k dipoles maximum. Homogeneous environment only (layered media and periodic structures listed as future work) arXiv:2505.09545v3. Ponomareva, Wiecha et al., SciPost Phys. Codebases 60 (2025), doi:10.21468/SciPostPhysCodeb.60
CPDDA (GL, paper) Not stated [paper], [GL] Paper 2025-03; repo last active 2025-04 Likely dormant Not on PyPI CuPy, with 3D-FFT matrix-vector products; tested on Windows 10 with cupy-cuda12x No Cross sections; near and far fields; arbitrary shapes About 26× GPU speed-up over CPU [paper]. The license is unclear, so it is unusable as a dependency, but it shows that CuPy FFT-DDA works on Windows. Xu et al., Nanomaterials 15, 500 (2025), doi:10.3390/nano15070500
dda-bench (GH) GPL-3 0.1.1, 2026-03-02 [PyPI] New any Benchmarks the CPU and GPU builds of the wrapped codes — A harness that runs DDSCAT, ADDA and IFDDA and compares their outputs digit by digit The DDA-only precedent for our benchmark; cite it. Argentin, Chaumet, Gross & Yurkin, CPC 325, 110172 (2026), arXiv:2603.02871
HoloPy (GH) GPL-3 3.5.0, 2021-05 [GH] (the PyPI holopy name belongs to an unrelated package) Low conda No No Wraps Mie, multisphere T-matrix and ADDA A historical multi-method wrapper (see §5). —

No JAX DDA was found by search. Every discoverable PyTorch or CuPy DDA is GPL or unlicensed.

Gap confirmed: an MIT/BSD, FFT-accelerated (O(N log N) matrix-vector product), iterative-solver DDA with: - GPU support through PyTorch (native Windows CUDA, Linux CUDA/ROCm) and a CPU path through NumPy/SciPy; - autodiff through implicit or adjoint differentiation of the linear solve, rather than unrolling the solver iterations; - modern polarizabilities (LDR, FCD, IGT) and substrate support.

4. FDTD

Library License Latest release Activity Install / Windows GPU Autodiff Dispersive metals Notes and limitations Citation
Meep (GH) GPL-2+ [GH] v1.35.0, 2026-10-09 [GH] Very active (1.8k stars) conda-forge only. No PyPI package (the PyPI meep and pymeep names are unrelated or 404). "Native Windows installation is currently unsupported"; use WSL doc. No (CPU, MPI) Adjoint solver (meep.adjoint, autograd/JAX wrappers) Yes (Lorentz/Drude susceptibilities) Subpixel smoothing; near-to-far field; Bloch periodic boundaries; the reference open FDTD. Optional plugin only, via a conda or WSL subprocess. Oskooi et al., CPC 181, 687 (2010)
FDTDX (GH, docs) MIT [GH] (PyPI license field empty) 0.6.3, 2026-10-09/12 Very active (369 stars) any, Python 3.11–3.14. Extras [cuda12], [rocm] and [tpu] (since 0.4.3). Native Windows is CPU only, because JAX CUDA is Linux-only. Depends on tidy3d (LGPL) [PyPI]. JAX: CUDA and ROCm (Linux), TPU, multi-GPU Yes, JAX AD with time-reversal memory saving [doc] Drude/Lorentz ADE merged 2026-05-13 (v0.6.2+) [GH commits]. The JOSS paper says non-dispersive only. Complex permittivity and conductivity added 2026-07. PEC, PMC and Bloch boundaries since v0.6.2; GDS import. No published plasmonic or Mie validation found. Mahlau et al., JOSS (2026), doi:10.21105/joss.08912
Tidy3D (GH) Client LGPL-2.1; solver proprietary and cloud-only 2.12.2, 2026-10-28 [PyPI] Very active any; uv OK on all OSes. Runs need an API key and paid FlexCredits; there is an Education Free tier of 10 credits/month (edu, billing). tidy3d-extras adds local functionality (its license field is empty). Cloud GPU Yes (autograd adjoint; autograd is a core dependency) Yes; documented gold-sphere-vs-Mie notebook (doc) Best-in-class accuracy and UX, but not local and not free. FlexAgent MCP server for AI agents (Nov 2025). Optional cloud backend only. —
TorchFDTD (GH) MIT [GH] 1.1.7, 2026-10-06/07 New (Sep 2026, 24 stars) any. Explicit native Windows + NVIDIA instructions using the PyTorch cu126 index and CuPy kernels [PyPI readme]. PyTorch + CuPy fused CUDA kernels; CPU fallback Yes (discrete adjoint into torch autograd) Unverified Browser CAD workbench; validated against Meep (per its docs). Very young. Park & Park, arXiv:2609.30039
fdtd (flaport, GH) MIT 0.3.5, 2023-12-07 Low (push 2025-09) any PyTorch torch.cuda backend (works on Windows) Not documented No Drude/Lorentz (conductivity and anisotropy only) [doc] An educational code; not suitable for plasmonics without major work. —
fdtdz (GH) MIT 1.1.3, 2023-10 Dormant (push 2025-02) sdist; JAX custom CUDA kernel; Linux CUDA JAX No (dielectric photonics) Specialised for thin-slab silicon photonics. —
ceviche (GH) MIT 0.1.3, 2023-02 Dormant any No HIPS autograd No 2D FDFD/FDTD; teaching and inverse design. Hughes et al., ACS Photonics (2019)
Khronos.jl (GH) MIT — Archived (2026) [GH] Julia KernelAbstractions: CUDA, ROCm, oneAPI, Metal Yes No polarizable materials (per a secondary listing) Dead end, but a good design reference. —
Luminescent.jl (GH) / luminescent (PyPI 4.3.1) No license file on GitHub; PyPI license empty (unverified, so treat as proprietary) 4.3.1, 2026-10-28 Active Julia core with a Python wrapper GPU Yes Unverified License unclear, so avoid it. —
gprMax 4 (GH) GPL-3+ 4.0.1, 2026-10-03 [PyPI] Active Wheels for Win, Linux and macOS (cp311–cp313) CUDA, OpenCL, Apple Metal (per the v4 announcement; partly unverified) No Debye/Drude/Lorentz (GPR-oriented) Built for ground-penetrating radar and antennas, not nano-optics. —
openEMS (GH) GPL-3 v0.37.0, 2026-10-09 Active Windows builds; Python interface via pyopenems on PyPI (third party?) No No Yes RF-oriented EC-FDTD. Not a nano-optics fit. —

Takeaway. - No permissive, pip-installable, native-Windows-GPU FDTD has been validated for plasmonics. - FDTDX is the strongest MIT GPU candidate (JAX, Linux GPU), but its dispersive support is brand new. - TorchFDTD is the only MIT FDTD with native Windows CUDA, but it is very young and its dispersion support is unverified. - Meep remains the CPU accuracy reference, but it is GPL and conda-only, so it can only be an optional plugin.

5. Multi-method / unified toolkits (potential competitors)

Project What it unifies How it differs from Lightsmith
HoloPy (GH), GPL-3 Mie (layered), multisphere T-matrix (SCSMFO) and ADDA behind one scatterer/theory API Holography and inference focus; no FDTD; no GPU or autodiff; last release 2021; conda-only; GPL.
TorchGDM (paper), GPL-3 GDM plus effective-dipole models from Mie and T-matrix (via treams) One method family; dense and limited N; homogeneous environment; GPL. The closest "multi-scale" competitor.
Tidy3D, LGPL client with proprietary cloud solver FDTD, mode solver, EME, heat/charge, RF Commercial cloud; no Mie, T-matrix or DDA solvers, though Mie is used for validation; has an agent MCP.
smuthi, MIT T-matrix (via NFM-DS for non-spherical particles) plus layered media One method; packaging problems.
PyMieSim, MIT Mie for spheres, core-shell and cylinders, with detectors and experiment sweeps Mie only.
dda-bench, GPL-3 DDSCAT, ADDA and IFDDA harness DDA only; a benchmark, not a solver API.
FermiLink (arXiv:2604.03460) LLM agent spanning about 50 simulation packages An agent layer over existing codes, not a unified physics spec.
ScattPort (scattport.org), Wikipedia list (DDA codes) Catalogues of scattering codes Directories only.

Conclusion. No project offers one declarative problem spec (particle + materials + environment + illumination + observables) dispatched to Mie, T-matrix, DDA and FDTD with consistent units, conventions and outputs, a validation harness, and agent-friendly schemas. That is the Lightsmith niche.

6. Refractive-index data sources (programmatic)

Source License Notes
refractiveindex.info database (GH) CC0-1.0 [GH] (public-domain dedication) YAML; actively curated (push 2026-10-27; snapshot release 2026-05-24). Cite Polyanskiy, Sci. Data 11, 94 (2024), doi:10.1038/s41597-023-02898-2, plus the original dataset papers listed in each file [doc]. Best choice: vendor or download a pinned snapshot.
refractiveindex (toftul, GH) MIT 1.3.1, 2026-10-27. Python interface to the database above.
PyOptik (GH) MIT 3.5.16, 2026-10-27. Unit-aware materials, Sellmeier and tabulated data, thin films.
refidx (GL) Conflict: pyproject.toml and PyPI say MIT, but GitLab's license detector reports GPL-3.0-only 1.3.0, 2025-10. Clarify the license before depending on it.
dispersion (JCMwave, GH) MIT 1.0.6, 2024. Database file handling.
Tidy3D material library and dispersion fitter LGPL client Useful for Drude-Lorentz pole fitting, which FDTD needs (fitter existence from memory; re-verify the API).
Meep materials library GPL-2+ Do not copy the code. The underlying literature fit parameters (e.g., Rakić 1998) are facts and can be cited and re-entered.

Recommendation. Ship a small MIT module that: - reads refractiveindex.info YAML (CC0); - stores the provenance and citation of each dataset; - provides analytic Drude-Lorentz / critical-point fits for Au, Ag, Al and Cu (e.g., Johnson & Christy, Rakić, Etchegoin), so that all four methods use identical ε(λ). FDTD requires a fit while the others use tabulated data, so the benchmark should report the fit error separately.

Handbook compilations such as Palik may carry publisher copyright on the compilation; prefer primary-paper datasets (unverified legal nuance).

7. Published cross-method benchmarks for plasmonic and small particles

Status key: - V = verified from the abstract, publisher page or arXiv page. - P = citation verified, but the takeaway comes from abstracts or snippets. - U = unverified.

Classic comparisons

  1. Wriedt & Comberg (1998), JQSRT 60, 411, doi:10.1016/S0022-4073(98)00016-8. P
  2. Compared DDA (DDSCAT), VIE, FDTD and EBCM on a dielectric cube.
  3. FDTD is general but costly; DDA is fast but memory-hungry.
  4. Kahnert (2003), JQSRT 79–80, 775, and the 2016 tutorial, JQSRT 178, 22. V
  5. Qualitative method reviews, not numerical benchmarks.
  6. Yurkin, Hoekstra, Brock & Lu (2007), Opt. Express 15, 17902, doi:10.1364/OE.15.017902. P
  7. DDA vs FDTD for dielectric spheres (x ≤ 80, m ≤ 2).
  8. DDA is faster at small m.
  9. Myroshnychenko et al. (2008), "Modelling the optical response of gold nanoparticles," Chem. Soc. Rev. 37, 1792, doi:10.1039/B711486A. P
  10. A tutorial review with a critical comparison of methods, including FDTD and BEM.
  11. The full method list and its findings are U.
  12. Hoffmann et al. (2009), Proc. SPIE 7390, 73900J, arXiv:0907.3570. V
  13. Compared FEM codes (JCMsuite, COMSOL, HFSS, CST) against MMP on gap-antenna near fields.
  14. Time-domain results largely did not arrive in reasonable time.
  15. Smajic et al. (2009), J. Comput. Theor. Nanosci. 6, 763. P
  16. Compared FEM and FDTD against MMP, MAS and BEM on 2D plasmonic wires.
  17. Boundary methods win in 2D, and FDTD shows staircasing errors.
  18. Yurkin, de Kanter & Hoekstra (2010), J. Nanophotonics 4, 041585, doi:10.1117/1.3335329. V
  19. ADDA on gold spheres, cubes and rods of 10–100 nm.
  20. Keeping near-IR absorption error under 10% needs about 10⁷ dipoles per sphere.
  21. Shape (staircase) error dominates, and cubes are about 10× better than spheres.
  22. FCD beats LDR.
  23. This is the key motivation for a fast GPU DDA and for reporting dipole-count convergence.
  24. Parsons, Burrows, Sambles & Barnes (2010), J. Mod. Opt. 57, 356, doi:10.1080/09500341003628702. P
  25. Compared Mie, null-field T-matrix, DDA, FEM and FDTD on a gold sphere.
  26. The closest classic analogue to our paper, but a one-off with no released code.
  27. Karamehmedović et al. (2011), Opt. Express 19, 8939. P
  28. Compared NFM-DS, COMSOL and MMP on Ag spheres and spheroids in the near field.
  29. Loke et al. (2014), Appl. Phys. B 115, 237, doi:10.1007/s00340-013-5594-z. P
    • Compared DDA (including DDA-SI with a substrate) against FDTD, FEM, Mie and NFM-DS on Au spheres, free-standing and on a substrate.
  30. Calà Lesina et al. (2015), Opt. Express 23, 10481, doi:10.1364/OE.23.010481. P
    • Checked FDTD for Au and Ag spheres against Mie.
    • Needs a mesh of about 0.25–0.5 nm for about 1% error, a Drude plus two critical-point dispersion model, and double precision.
  31. Gallinet, Butet & Martin (2015), Laser Photonics Rev. 9, 577, doi:10.1002/lpor.201500122. V (abstract only)
    • Review of "standard problems" in nanophotonics.

Recent work (2019–2026)

  1. Lalanne et al. (2019), JOSA A 36, 686, arXiv:1811.11751. V
    • A multi-group benchmark of quasinormal-mode solvers, written to set standards.
    • A model for our paper's framing.
  2. Rassinfosse et al. (2024), Optics Continuum 3(6), doi:10.1364/OPTCON.511183. P
    • FDTD on periodic plasmonics.
    • "No general rule" for mesh size exists.
  3. Dieperink et al. (2024), Nanophotonics 13, 4647, doi:10.1515/nanoph-2024-0238, arXiv:2410.08093. P
    • Compared DDA, FDTD, DGTD and BEM against Mie for a 50 nm Au sphere.
    • BEM was fastest by orders of magnitude, and DDA showed irregular convergence.
    • The closest recent prior work, but a single case with no harness.
  4. Argentin, Chaumet, Gross & Yurkin (2026), CPC 325, 110172, arXiv:2603.02871, code: dda-bench. V
    • DDSCAT, ADDA and IFDDA agree to 10–15 digits when hidden parameters are aligned, and only to about 4 digits otherwise.
    • Fair benchmarking is hard because of defaults and stopping criteria.
    • Covers CPU and GPU energy use.
    • Test cases are dielectric only (no metals, no Mie reference).
  5. Tidy3D gold sphere vs Mie notebook. V, but vendor material and not peer-reviewed.
    • Notes that subpixel averaging is off for metals.

Novelty check. No open, scripted, reproducible benchmark was found that runs Mie, T-matrix, DDA and FDTD open-source codes against Mie for plasmonic particles and reports both accuracy and cost. - The precedents are each narrower: #16 is DDA-only and dielectric, #15 is a single case, and #8 and #10 are one-offs. - Before submission: - search JQSRT 2025–26 and the ELS 2026 proceedings; - check the validation sections of the smuthi, treams and MNPBEM papers; - check anything citing arXiv:2603.02871. - Methodological lesson from #16: expose and align the hidden parameters (polarizability prescription, iterative tolerance, stopping rule, precision), and report them.

8. LLM agents running photonics/EM simulations (2023–2026)

Work Simulator What the agent does Lessons Status
MetaChat: Lupoiu … Fan, Sci. Adv. (Oct 2025), doi:10.1126/sciadv.adx8006, arXiv:2503.20479 DL surrogate Maxwell solvers Multi-agent freeform metasurface design Agents work with fast surrogates in the loop. V
Kim, Park & Shin, Nanophotonics (2025), doi:10.1515/nanoph-2024-0674 Transfer matrix (LLM-written code) In-context design of multilayers and metasurfaces — V
Huang et al., MCP-enabled LLM for meta-optics inverse design, arXiv:2508.10277 TorchRDIT (differentiable RCWA) via an MCP server with 5 APIs Inverse design Structured prompts and a small API surface beat free-form natural language on quality, cost and errors. V
PhIDO: Sharma … Poon, arXiv:2508.14123 PIC layout and simulation Natural language to PIC layout Up to 91% success on single devices; about 57% pass@5 end-to-end. V
PICBench, arXiv:2502.03159 Circuit simulator Netlist generation Iterative error feedback matters. V
PICopilot, arXiv:2608.01791 Unstated Multi-agent retrieval-augmented script generation Solved 48/48 tasks vs 27/48 for a GPT-5 baseline with generic retrieval. V
Kharel, Khavasi, Chen & Hughes, arXiv:2606.00915 Unstated (Tidy3D likely, U) Propose, simulate, evaluate against acceptance criteria — V
HALO with HALO-Bench (52 tasks), arXiv:2608.28877 Nanophotonic solvers (unstated) Compares a fixed workflow, a structured agent and a coding agent Typed specs plus a fixed workflow are the most reliable and token-efficient; free-form coding adds operational failures. V
Biswas, De & Islam, arXiv:2606.22493 Eigenmode solver + FDTD (MPB/Meep, U) Directional-coupler design with convergence judgment — V
Flexcompute FlexAgent MCP (news, docs) Tidy3D Builds and modifies simulations, troubleshoots validation errors; works with Claude Code, Cursor and Codex Vendor product with no metrics. Tidy3D now hard-depends on fastmcp [PyPI]. V
FEABench, arXiv:2504.06260 COMSOL API LLMs generate FEA models Executable 88% of the time, but no model fully solved any problem: "runs" is not "correct". P
FermiLink, arXiv:2604.03460 About 50 packages (a Meep variant, U) Reproduces published figures 74/132 figures reproduced (56%). V

Design implications for Lightsmith: - a Pydantic/JSON-Schema problem spec with a small typed API surface; - an MCP server exposing schema, validate, run and compare; - machine-readable errors with fix hints; - built-in physics checks (an optical-theorem check, Mie agreement, a convergence flag), so that an agent can tell "ran" from "correct".

No agent or benchmark targets single-particle scattering codes with an analytic ground truth, which is a differentiator. The search was not exhaustive.

9. GPU array-framework realities on native Windows (late 2026)

Framework Native Windows GPU Linux GPU macOS GPU Autodiff FFT / complex128 Install notes
PyTorch 2.15.0 (2026-10-28) Yes, NVIDIA CUDA. Wheels for cu126, cu130, cu132 and cu134 (win_amd64) are on download.pytorch.org (checked index). AMD ROCm on Windows is a preview through AMD's own wheels (ROCm 7.2, Jan 2026; RX 7000/9000 and Ryzen AI only) (AMD notes). CUDA and ROCm MPS. No float64 on MPS (widely reported, e.g. Apple forum); complex support partial and improving (2.15 adds complex64 LU on MPS, from secondary source). Yes torch.fft with complex128 on CUDA and CPU The PyPI torch Windows wheel is CPU-only. uv needs a [[tool.uv.index]] entry for the PyTorch CUDA index with explicit = true plus tool.uv.sources.
JAX 0.12.0.1 (2026-10-21) No (official). Native Windows is CPU-only ("experimental"). jax-cuda12-plugin and jax-cuda13-plugin ship Linux wheels only [PyPI]. The docs say pip CUDA installs "do not work with Windows" (docs). WSL2 + NVIDIA is "experimental". CUDA 12 and 13; ROCm Not supported; jax-metal last released 2024-10 [PyPI] Yes (best-in-class) Yes Unofficial winjax 0.12.0 (GH, beta, 0 stars) loads a Windows-built PJRT CUDA plugin. Single GPU; cannot coexist with PyTorch in one process; no Pallas. Not a foundation to build on.
CuPy 14.2.0 Yes. cupy-cuda12x and cupy-cuda13x Windows wheels; the [ctk] extra pulls the CUDA toolkit from PyPI (docs) CUDA; ROCm experimental (no official ROCm wheels from v13.4) No No cuFFT, complex128 Good for a fast non-differentiable path. Zero-copy with torch via DLPack.
NVIDIA Warp 1.18.0 Windows wheels [PyPI] CUDA macOS-arm wheels (CPU) Yes (kernel-level) Limited A kernel DSL, a poor fit for FFT-heavy DDA.
nvmath-python 1.0.0 Windows wheels Linux — No cuFFT, cuBLAS with device-side callbacks A possible accelerator under CuPy or torch.

Implications for the GPU strategy: 1. Use PyTorch as the primary GPU and autodiff backend for our own solvers (DDA first, then the differentiable Mie). It is the only option that gives native Windows CUDA, Linux CUDA/ROCm, a CPU fallback, and autodiff in one install. Keep a pure NumPy/SciPy CPU reference path, and consider the Python Array API standard (array-api-compat) to share code. 2. Treat JAX as a "Linux/WSL2 backend family" that reaches FDTDX. Run it out of process (subprocess or worker), so that JAX and torch never share a process, which also sidesteps the winjax conflict. 3. FP64 is a real concern. Plasmonic accuracy studies call for double precision (#11), and consumer GeForce GPUs have very low FP64 throughput. Benchmark complex64 vs complex128 (and mixed-precision iterative refinement) explicitly, as both a paper result and a design decision. 4. On macOS, MPS has no float64, so default to the CPU there.

10. Recommendations

Per-method backend choices (MIT-core-compatible first; GPL = optional plugin only)

Method Best CPU backend (pip/uv, permissive) Best GPU backend (pip/uv, permissive) Optional plugins (GPL or other caveats) Our own work
Mie miepython (MIT) for homogeneous spheres; treams (MIT) for layered spheres and T-matrix-consistent Mie None needed for speed. For autodiff and batched GPU, write our own MIT PyTorch Mie (a few hundred lines; log-derivative recurrences). scattnlay (GPL; multiprecision cross-check), PyMieDiff (GPL), PyMieScatt (MIT but stale) A small, high-value MIT differentiable Mie; also the benchmark ground truth. Audit it against scattnlay's multiprecision mode.
T-matrix treams (MIT; Windows, Linux and macOS-arm wheels; clusters, lattices, stratified media) None exists that is pip-installable (CELES is MATLAB) smuthi (MIT itself, but has a GPL pywigxjpf dependency and numpy<2), pyMSTM / MSTM (MIT, Linux wheels only), TERMS (MPL-2.0 executable), MiePy (GPL). Exclude SMARTIES (CC BY-NC). Upstream work (§ below). A GPU T-matrix is a lower-priority gap. Read and write the KIT T-matrix HDF5 format.
DDA / GDM No permissive pip-installable option exists. Use our own MIT FFT-DDA (NumPy/SciPy, CPU) Our own MIT FFT-DDA in PyTorch (CUDA on Windows and Linux, ROCm on Linux) ADDA (GPL-3 executable via subprocess; the accuracy reference, with OpenCL GPU), pyGDM2 and TorchGDM (GPL-3), DDSCAT (GPL) Highest-value gap (see below).
FDTD No permissive CPU code is validated for plasmonics. Short term, use FDTDX on CPU (MIT; runs on native Windows CPU, slowly). FDTDX (MIT, JAX; Linux/WSL2 CUDA or ROCm). TorchFDTD (MIT; native Windows CUDA) is a watch/evaluate candidate. Meep (GPL-2+, conda/WSL, CPU reference); Tidy3D (LGPL client + paid cloud, "cloud plugin"); flaport/fdtd (MIT but no dispersion) Contribute plasmonic validation, and fixes if needed, to FDTDX's new Drude-Lorentz ADE. Our benchmark is that validation.

The most valuable gap to fill ourselves: a permissive GPU + autodiff FFT-DDA

Today's options are GPL (ADDA, DDSCAT, pyGDM2, TorchGDM), unlicensed (CPDDA), or dense and size-limited (TorchGDM, about 10–15k dipoles). Yet the literature (#7) says plasmonic accuracy needs 10⁶–10⁷ dipoles, which only FFT plus an iterative solver on a GPU makes interactive.

Minimum viable feature set: - Polarizabilities: CM, LDR, FCD and IGT. Use FCD as the default for metals, per #7. - FFT-accelerated interaction matvec in torch with complex64 and complex128. - Krylov solvers: BiCGStab, QMR and GMRES. ADDA's defaults (QMR) are a good reference. - Gradients via implicit differentiation (an adjoint solve with the transposed system) for memory-safe autodiff. - Observables: C_ext, C_abs and C_sca; near and far fields; orientation averaging. - Later: substrate via a Sommerfeld or image approximation, and periodic structures. - Validate against ADDA bit-for-bit where possible, reusing dda-bench's aligned-parameter methodology, and against Mie.

Upstream contributions (cheap, high goodwill, and they unblock our backends)

  • smuthi:
  • release Meson-built wheels for cp311–cp313 on Windows and macOS;
  • lift the numpy<2 and scipy<1.14 pins;
  • make pywigxjpf (GPL) optional or replace it with a permissive Wigner-3j implementation. This also removes the license wrinkle.
  • pytmatrix: a Meson/f2py port so it builds on Python ≥3.12.
  • scattnlay: publish 2.5 to PyPI with cibuildwheel for Linux and more Python versions.
  • pyMSTM: Windows and macOS wheels.
  • treams: Python 3.14 support and relaxing the scipy cap.
  • FDTDX: plasmonic test cases (Au and Ag spheres vs Mie, with a Drude-Lorentz fit); declare the license in the PyPI metadata (the field is currently empty).

Packaging policy for the MIT core

  • Core dependencies: numpy, scipy, pydantic, miepython, treams. These are all MIT/BSD with Windows wheels.
  • Extras:
  • [torch] for the GPU DDA and Mie;
  • [jax] for FDTDX, Linux-recommended and run out-of-process;
  • [tidy3d] for the cloud backend;
  • [gpl] meta-extra for scattnlay, pyGDM2 and TorchGDM, documented as changing the license of the combined work if redistributed.
  • ADDA and Meep are discovered as external executables or environments, never imported.
  • Avoid depending on packages with exact version pins (PyMieSim) or GPL transitive dependencies (smuthi) in the core.

Main sources not linked inline: PyPI JSON API pages https://pypi.org/pypi/<name>/json for every package named above, and the GitHub and GitLab REST APIs, all queried 2026-11-01.